Glossary
the terms a buyer needs before writing a board paper
Terms that decide a project
Not a dictionary of AI. These are the terms whose definition changes what you buy.
Every entry ends in a comparison, because a definition on its own rarely settles anything - what settles it is knowing what the term is being contrasted with, and which of the two your situation actually calls for.
Semantic search
Retrieval that matches meaning rather than words - which matters because the asker does not know the document's vocabulary.
Document extraction accuracy
A single accuracy figure for document extraction is close to meaningless. The split by input quality is the real number.
Full coverage
Reviewing everything rather than a sample. The gain is rarely speed - it is finding what the unsampled remainder was hiding.
Three-way match
Reconciling the purchase order, the delivery note and the invoice. Simple to state, and the single most common place manual document work hides.
Human in the loop
A designed review path for the cases the system is not confident about - not a fallback, and not an apology for the model.
AI readiness assessment
Scoring whether an organisation can absorb a specific AI deployment. Readiness for something, never readiness in the abstract.
Retrieval-augmented generation
Answering from your documents rather than from the model's memory. The retrieval half is where it succeeds or fails.
Bilingual document processing
Handling Arabic and English natively rather than translating first - which is where most global tooling quietly degrades.
Data readiness
Whether the data can support the decision you want to automate. Usually the binding constraint, and usually not what the client came in asking about.
Workflow copilot
An assistant inside the tools people already use. The value is in not being a fifth place to go and ask.
Voice agent
A system that answers the phone and completes the transaction. Materially harder than a chat widget, and the failure modes are specific.
Model drift
A system that was accurate at handover quietly becoming less so as the world it was measured against changes.
Sovereign deployment
Running the system where the data is legally required to stay. Table stakes in this region rather than a differentiator.
Audit before automate
Establishing what a process costs and whether it is sound before building anything on top of it.
Problem shape
The structural form of a problem, independent of industry. It is what makes proof from one sector usable in another.
Agent-based automation
Usually sold as agentic AI: a system that takes a sequence of actions towards a goal rather than answering one question at a time. The useful question is never whether it is agentic. It is what it is allowed to do without asking.
Hallucination
A fluent, confident answer that is not supported by any source. The problem is not that it is wrong. It is that it is indistinguishable from a right answer.
OCR and document AI
OCR turns pixels into characters. Document AI decides what those characters mean, checks them against what the business already knows, and flags what disagrees. Buying the first and expecting the second is the most common disappointment in this category.
AI governance
The named answers to who approved this system, who is accountable when it is wrong, what it is allowed to touch, and how anyone would notice it degrading. Not a document. Four answers with names against them.
Build versus buy
The decision is almost never about whether your team could build it. It is about what else they would not be doing for the six to twelve months it takes them to learn what somebody else already knows.
Start here
If a term here is the one your board paper turns on, ask us about it.
We will send the entry, the evidence behind it, and the honest note about where it does not apply.
You get a reply within one working day, from the engineer who would do the work - not a sales sequence.